The questions people actually ask.

Built-in memory belongs to the vendor; Smara belongs to you. What ChatGPT remembers, Claude doesn't — and none of it is yours to inspect, correct, or take with you. Smara is one memory that every assistant you connect shares: tell one thing to Claude and ChatGPT knows it too.

It is also a different kind of memory. Built-in memory is a flat list of notes. Smara consolidates what you say into durable facts with provenance, a knowledge graph of the people and things in your life, and an append-only ledger of where every fact came from — all of it visible in the portal, all of it exportable.

Anything that speaks MCP — which today means practically everything. Claude and Grok connect with just the server URL (https://api.smara.dev/mcp); ChatGPT and Gemini need one OAuth credential you mint in the portal. Developer tools like Claude Code and Cursor connect with a single config block, and your own agents can use the same MCP endpoint or the plain REST API underneath.

The same key and the same memory back every surface, so there is nothing to sync.

At first, yes — assistants only call connected tools when the conversation gives them a reason to. The reliable opener is "Use Smara to remember that…" or "Check my Smara memory — what do you know about me?".

To make recall automatic in Claude, install the smara-memory Skill from the docs: it teaches Claude to check memory at the start of every conversation and record durable facts as they appear, without being asked. On other assistants, pasting the short memory prompt from the docs into custom instructions does the same job.

Four kinds of memory, separated on purpose. Short-term working memory (recent turns, scratchpads) expires on its own within hours to days. Episodic memory is the append-only ledger of what happened. Semantic memory is durable facts distilled from those events, each with confidence, validity dates, and links back to the sessions it came from. And the knowledge graph holds the entities — people, projects, places — and how they relate.

Nothing is stored that wasn't said to a connected assistant, written through the API, or imported from an export you uploaded. All of it is browsable in the portal.

Every night Smara consolidates — we call it dreaming. Contradictions get resolved, duplicates merge, related facts form communities, and the knowledge graph is reorganized. You wake to a dream report of what changed.

Unlike other systems that quietly rewrite themselves, every edit a dream makes is recorded in the audit trail. You can open any morning's report and see exactly what was merged, superseded, or reorganized — and why.

Your memory is yours alone. Every record carries its tenant, and cross-tenant reads are not a configuration option — they are structurally impossible in the data model. Nobody else's assistant, and no other customer, can reach your memory.

Data lives on Google Cloud in the United States. Memory content is processed by Google's Vertex AI services during consolidation and search under terms that do not permit Google to train on it. We do not sell data and we do not use memory data for advertising.

No. We do not use your memory data to train our own models or anyone else's, and our cloud subprocessors are bound by terms that forbid them from training on it too. Your memory exists to serve you back your own context — that is the entire business.

Two ways. In conversation, just tell your assistant — "actually, I moved to Seattle" — and the old fact is superseded by the new one, with the correction recorded in the audit trail. In the portal, open Memories, find the fact, and edit or forget it directly.

Forgetting is real deletion: the fact stops being retrievable immediately and is permanently removed after a retention window. Memory also forgets on its own — facts that stop mattering decay and eventually expire.

Yes, all of it, any time. GET /v1/export streams everything back out: facts, the knowledge graph, episodic events, the full revision audit trail, your import registry, pools, and skills. There is no export fee, no request form, and no degraded "data dump" — it is the same structured data the service runs on.

Anti-lock-in is the founding argument of the product: memory that can't leave isn't yours.

Yes — this is the fastest way to make Smara useful on day one. Download the official data export from each provider, run the smara-import CLI (a zero-dependency Python script), and consolidation distills years of conversations into durable facts dated when they actually happened.

Imports are idempotent: re-uploading the same export is a no-op, and a fresher export ingests only what is new. Re-run it as often as you like.

Smara is free during the preview. Sign up, connect your assistants, import your history — no card required. Paid plans will come with general availability, and preview users will get generous notice and a founding-user price before anything changes. Your export rights are the same either way.

No SDK, nothing to install. Smara is a hosted MCP server and a plain REST API on the same handlers, so the two surfaces cannot drift apart. A working integration is three REST calls — read context in, write events out, search what you know — or one MCP config block if your framework speaks MCP.

Multi-agent teams get two extra primitives: shared memory pools with per-key grants, and signed webhooks that fire when memory changes.